One agent. Coherent across sessions.
Atlas remembers Priya's renewal terms across every call. Cleo carries last quarter's escalation pattern into this morning's ticket. No re-priming, no cold starts.
Sonzai · Multiplayer MemoryMind Layer for teams of agents
One memory.
Every agent. Every teammate.
A typed knowledge graph shared by every AI Employee on your project — and every human reading alongside them. Documents ingest, agents write back, newer facts supersede older ones. The operating context for a company OS.
Live above · agents · users · documents · facts
What it isSix things that make it
Six things that make it
a shared memory, not a vector store.
A typed graph. Source-anchored. Time-aware. Written by agents, read by agents and humans. Conflict resolution baked in.
Every agent. The same graph.
Sales updates a milestone. Support sees it on the next ticket. Ops plans capacity against it. Knowledge agent enriches it. The same typed facts, all the way down.
Typed facts, not chat history.
Person, Account, Decision, Promise, Document, Fact. Source-anchored, timestamped, with confidence. Retrieval grounds on what actually matters — not yesterday's transcript fragments.
Drop a doc. Watch the graph absorb it.
PDFs, Notion exports, Google Docs, Slack threads, call transcripts. Sonzai extracts typed facts, anchors them to the source, and slots them into the graph. Every agent picks up the context on their next turn.
Agents write back what they learn.
When Atlas closes a discovery call, the model emits new facts. The runtime persists them with audit trail and CAS update semantics. Concurrent agent writes don't collide — every change is durable.
Newer information wins.
Every fact carries time, source, and confidence. When new evidence conflicts with old, Sonzai resolves on recency + source priority — newer wins, old is archived for audit, stale facts decay in retrieval weight.
Where it landsThree shapes of value,
Three shapes of value,
on the same runtime.
Company OS
A live operating picture for the whole org.
Decisions, promises, accounts, processes — all typed, all queryable, all current. Humans browse it, agents ground on it, new docs flow in. The institutional memory that doesn't leave when employees do.
AI Employees in a team
Multiple agents, one operating context.
Sales, support, ops, knowledge — each a focused agent, all sharing the same memory graph. The handoff between them is instant because there's no handoff: they're all looking at the same picture.
Customer continuity
The customer never has to re-explain.
Whatever they told sales, the support agent already knows. Whatever they told support last quarter, the renewal agent has in mind. Across reps, channels, and months — one memory of the relationship.
Shape of the APITwo agents.
Two agents.
One memory.
Sales agent records a Decision. Support agent reads it on the next conversation. The runtime handles typing, source anchoring, conflict resolution, and recency decay.
- —knowledgeBase: true — read from project KB
- —knowledgeBaseWrite: true — agent writes back what it learns
- —knowledgeBaseScopeMode: cascade — project + tenant-wide org KB
- —Built-in CAS update semantics so concurrent writes never clobber
// Sales agent commits a fact. await client.agents.chat({ agent: atlas.agentId, userId: "acme-priya", knowledgeBase: true, knowledgeBaseWrite: true, // Atlas writes back knowledgeBaseScopeMode: "cascade", messages: [ { role: "user", content: "We agreed Q3 renewal at $40k." } ], }); // Hours later. Different agent, same KB. const reply = await client.agents.chat({ agent: cleo.agentId, // Support agent userId: "acme-priya", knowledgeBase: true, // Cleo reads what Atlas wrote messages: [ { role: "user", content: "What was our renewal commitment?" } ], }); // Cleo: "Q3 renewal at $40k, agreed with Atlas on May 7."
The institutional memory
that doesn't leave.
Spin up the platform yourself, or have us design and build the first AI Employees on top of your knowledge base. Either way, the memory compounds.